Flooding detection in the distillation column by ML tools

In the following sections, the preprocessing of the acquired sensor data, feature extraction, training of the ML methods and the implementation with live data are described.

Preprocessing of time series data

Multivariate time series data can be tricky to deal with as the temporal structure should be preserved in some way during the training process. One way to make supervised learning methods applicable to time series data is the sliding window method [40], which transforms data in such a way that past and “future” measurements are preserved for each data point. For this use case, the future pressure drop (\(p_{t+1}\),\(p_{t+2}\), …) will be predicted based on the past data of pressure drop and other significant parameters \(X_{i}\) (…,\(p_{t-1}\), \(X_{1,\ \ t-1}\), … , \(p_{t}\), \(X_{1,t}\), …). These other significant parameters are determined in section 3.2. A schematic representation of the sliding window data transformation is given in Figure 3. The window size refers to the time window of past data and the response size describes the forecast window.